similar to: PGO information at LTO/thinLTO link step

Hi Teresa, Actually, enabling the new pass manager manually seems to have solved this issue, so this problem is only valid for the old pass manager. Thanks, Graham Yiu LLVM Compiler Development IBM Toronto Software Lab Office: (905) 413-4077 C2-707/8200/Markham Email: gyiu at ca.ibm.com From: Teresa Johnson <tejohnson at google.com> To: Graham Yiu <gyiu at ca.ibm.com> Cc:

On Tue, Oct 3, 2017 at 1:46 PM, Teresa Johnson via llvm-dev < llvm-dev at lists.llvm.org> wrote: > > > On Tue, Oct 3, 2017 at 1:38 PM, Graham Yiu <gyiu at ca.ibm.com> wrote: > >> Hi Teresa, >> >> Actually, enabling the new pass manager manually seems to have solved >> this issue, so this problem is only valid for the old pass manager. >> >

Thanks Easwaran. This is what we've observed as well, where the old PM inliner was only looking hot/cold callee information, which have signficantly smaller boosts/penalties compared to callsite information. Teresa, do you know if there is some documentation/video/presentation on how PGO information is represented in LLVM and what information is passed via the IR? I'm finding some

Hi Graham, Thanks for sharing this. Are you planning on enabling the pass only on PGO? Even in non-PGO, I noticed some performance gains when we are aggressive in partially inlining the early return part, especially when the callee spill CSRs in the entry block. At a high level, I have two questions: 1. What is the main obstacle that prevent the pass from being enabled by default? 2.

Hi Graham, Thank you for offering help. I am trying to create a reproducer. The problem is that the crashes happen whilst LTO is used. One thing I am sure about IR is broken at compile time. Thanks, Evgeny From: Graham Yiu <gyiu at ca.ibm.com> Date: Friday, 10 November 2017 at 16:09 To: Evgeny Astigeevich <Evgeny.Astigeevich at arm.com> Cc: "junbuml at codeaurora.org"

Forgot to add that all experiments were done with '-O3 -m64 -fexperimental-new-pass-manager'. Graham Yiu LLVM Compiler Development IBM Toronto Software Lab Office: (905) 413-4077 C2-707/8200/Markham Email: gyiu at ca.ibm.com From: Graham Yiu/Toronto/IBM To: llvm-dev at lists.llvm.org Cc: junbuml at codeaurora.org, xinliangli at gmail.com Date: 11/02/2017 05:26 PM Subject: [RFC]

Hi David, The only reason I can see to use the 'pattern matching' part as a fall-back is in case we cannot inline the (what I'm assuming would be) a much bigger hot-path-only cloned function for whatever reason. What I'm assuming here is that after cold-region outlining, we may still have a large portion of the original function body to attempt to inline, whereas the pattern

Hi, I noticed some performance gains in some spec benchmarks without significant code size bloat when aggressively performing partial inlining, especially when the original callee spill CSRs in the entry block. I guess the partial inlining is not enabled mainly due to the code size. Is there any other issue which prevent the pass from being enabled? Do we have any plan or any on-going works

On Tue, Oct 3, 2017 at 12:08 PM, Davide Italiano <davide at freebsd.org> wrote: > On Tue, Oct 3, 2017 at 11:57 AM, Graham Yiu via llvm-dev > <llvm-dev at lists.llvm.org> wrote: >> Hello, >> >> I recently noticed that the new pass manager was not enabled at regular/thin >> LTO link step even if '-fexperimental-new-pass-manager' was specified in the

I second the fact that a way to outline specific function regions independently of the partial inliner sound very useful. I am not sure however if we would want a mode within the partialInliner or something completely independent. As a general question, does anybody has a clear idea of what are the constraints on the region CodeExtractor is currently able to handle ? Going through the code, it

Hi Jessica, Thanks for the feedback. I believe the existing partial inliner pass does use some common utilities like the code extractor to do outlining. Is that what you're referring to? I don't recall seeing any other passes that has outlining other than the machine outliner, but I may have missed something. I briefly looked at River's RFC and it seems he's mainly utilizing

Hi Evgeny, I just realized that if these are compile-time errors I can help investigate on my end. Do you have something I can use to reproduce? Cheers, Graham Yiu LLVM Compiler Development IBM Toronto Software Lab Office: (905) 413-4077 C2-707/8200/Markham Email: gyiu at ca.ibm.com From: Graham Yiu/Toronto/IBM To: Evgeny Astigeevich <Evgeny.Astigeevich at arm.com> Cc:

Hello, My team and I are looking to do some enhancements in the partial inliner in opt. Would appreciate any feedback that folks might have. # Partial Inlining in LLVM opt ## Summary ### Background Currently, the partial inliner searches the first few blocks of the callee and looks for a branch to the return block (ie. early return). If found, it attempts to outline the rest of the

On Tue, Aug 15, 2017 at 4:14 PM, River Riddle via llvm-dev < llvm-dev at lists.llvm.org> wrote: > Hey Graham, > I worked on pretty much this exact thing last year. I did something > similar to what you described, I traversed the CFG and built potentially > profitable regions from any given valid start node. At that point there > were several road blocks that prevented it

Hi David, So I've began doing some implementation on the outlining portion of the code. Currently, I got the partial inliner to outline cold regions (single entry, single exit) of the code, based solely on the existence of ProfileSummaryInfo (ie. profiling data). However, I have some concerns on how this will co-exist with the existing code that peels early returns. The control flow looks

Hello, I recently noticed that the new pass manager was not enabled at regular/thin LTO link step even if '-fexperimental-new-pass-manager' was specified in the compile step and link step commands. Upon closer inspection, it seems there's so real way to invoke the new pass manager path ('runNewPMPasses' in lib/LTO/LTOBackend.cpp) during link step. 'Conf.UseNewPM' is

On Fri, Sep 2, 2016 at 3:30 PM, Xinliang David Li <davidxl at google.com> wrote: > On Fri, Sep 2, 2016 at 3:16 PM, Piotr Padlewski > <piotr.padlewski at gmail.com> wrote: > > > > > > 2016-09-02 15:04 GMT-07:00 Xinliang David Li <davidxl at google.com>: > >> > >> On Fri, Sep 2, 2016 at 2:58 PM, Piotr Padlewski > >>

> On Mar 24, 2015, at 12:08 PM, Xinliang David Li <davidxl at google.com> wrote: > > On Tue, Mar 24, 2015 at 10:54 AM, Bob Wilson <bob.wilson at apple.com> wrote: >> >>> On Mar 24, 2015, at 10:53 AM, Diego Novillo <dnovillo at google.com> wrote: >>> >>> On Tue, Mar 24, 2015 at 1:48 PM, Bob Wilson <bob.wilson at apple.com> wrote:

The profile summary is saved in the global metadata ASAIK. If we want to calculate if something is hot/cold while choosing functions for importing, we would either need to read whole Module (which we clearly don't want to do) or duplicate this information in the summary, so we could get it without reading Module. 2016-09-02 15:49 GMT-07:00 Mehdi Amini <mehdi.amini at apple.com>: >

Functions in LLVM IR have a function_entry_count metadata that is attached in PGO compilation. By using the entry count together with the block frequency info, the compiler computes the profile count of call instructions based on which the hotness/coldness of callsites can be determined. Experiments have shown that using a higher threshold for hot callsites results in improved runtime performance